HBaseCon 2012 | Base Metrics: What They Mean to You - Cloudera

HBase Metrics: What they mean to you

David S. Wang
dsw@cloudera.com
5/22/12

HBaseCon 2012. 5/22/12
Copyright 2012 Cloudera Inc. All rights reserved

Agenda

• Motivation
• Where to get metrics
• Collection methods
• Metrics:
• HBase Operations
• Memory-related
• Latency-related
• Host-level
• Takeaways

HBaseCon 2012. 5/22/12 2

Motivation

• To understand the steady-state
behavior of your HBase cluster
• To debug when something bad
happens
• To help evaluate changes
• To figure out when you need
to buy more hardware

HBaseCon 2012. 5/22/12 3

Where to get metrics
JVM JVM
Region server
Master
Region
Region

JVM

HDFS
Metrics can be obtained DataNode
from all of these areas

Host
CPUs Memory Disks Network

HBaseCon 2012. 5/22/12 4

Collection methods

• Dump to file
• web UI /jmx
• jconsole
• Ganglia
• Cloudera Manager
• Many other tools out
there – suggestions
welcome

HBaseCon 2012. 5/22/12 5

HBase Operations: compactions

• hbase.regionserver.compactionQueueSize – size of queue of compactions
waiting to be processed
• h.r.compactionSize_avg_time – average bytes processed per compaction
• h.r.compactionTime_avg_time – average time per compaction in
milliseconds
• h.r.compactionSize_num_ops, h.r.compactionTime_num_ops – total
number of compactions so far
• TIP: Spikes in h.r.compactionQueueSize could mean all of your regions
are growing at the same rate and need to split/compact at around the
same time – time to presplit your regions or turn off auto-compactions
HBaseCon 2012. 5/22/12 6

HBase Operations: flushes

• h.r.flushQueueSize – size of queue of flushes waiting to be
processed
• h.r.flushSize_avg_time – average bytes processed per flush
• h.r.flushTime_avg_time – average time per flush in milliseconds
• h.r.flushSize_num_ops, h.r.flushTime_num_ops – total number of
flushes so far
• TIP: Small h.r.flushSize_avg_time and large flushQueueSize mean
premature flushes: may indicate that you need more RAM, or you
are ingesting and flushing faster than your disks can handle
HBaseCon 2012. 5/22/12 7

HBase Operations: region splits

• h.r.regionSplitFailureCount – number of unsuccessful splits
• h.r.regionSplitSuccessCount – number of successful splits
• TIP: Spikes in h.r.regionSplitSuccessCount may mean that you are
spending too much time splitting – think about pre-splitting your
regions
• TIP: Sustained high rates of h.r.regionSplitSuccessCount may mean
that you are ingesting fast enough to exceed your region size
configuration – change the configuration and/or presplit/disable
automatic splits
HBaseCon 2012. 5/22/12 8

HBase Operations: stores and storefiles

• h.r.stores – total number of stores
• h.r.storefiles – total number of storefiles
• TIP: Large ratio between number of h.r.stores and
h.r.storefiles may indicate it’s time to compact more
frequently – otherwise read performance can suffer

HBaseCon 2012. 5/22/12 9

Memory: memstores

• h.r.memstoreSizeMB – sum of all the memstore sizes in megabytes
• h.r.numPutsWithoutWAL – number of puts with
setWriteToWAL(false)
• h.r.mbInMemoryWithoutWAL – amount of data from puts with
setWriteToWAL(false) in megabytes
• TIP: Anytime you have non-zero values in h.r.numPutsWithoutWAL
and h.r.mbInMemoryWithoutWAL, you risk data loss if the RS
crashes, until the next flush when they go back down to zero.
Useful to detect applications that setWriteToWAL(false).

HBaseCon 2012. 5/22/12 10

Memory: block cache

• h.r.blockCacheCount - number of blocks in block cache
• h.r.blockCacheEvictedCount – number of blocks evicted in the last period
• h.r.blockCacheFree, h.r.blockCacheSize – free/occupied space in block
cache in bytes
• TIP: High sustained levels of h.r.blockCacheEvictedCount mean your block
cache is turning over due to heap size constraints – possibly more GCs
• Get more memory
• Repartition heap amongst the various processes on the host – more for RS

HBaseCon 2012. 5/22/12 11

Memory: block cache

• h.r.blockCacheHitCount, h.r.blockCacheMissCount – number of cache
hits/misses
• h.r.blockCacheHitRatio/h.r.blockCacheHitRatioPastNPeriods – hit ratio (cache
hits/total requests) for past period, past N periods
• h.r.blockCacheHitCachingRatio/h.r.blockCacheHitCachingRatioPastNPeriods
– hit caching ratio (cache hits from requests set to use block cache/total
requests set to use block cache) for past period, past N periods
• TIP: With sustained traffic, in general you want your block cache to be fully
utilized for maximum performance – high h.r.blockCacheHitRatio and
h.r.blockCacheHitCachingRatio

HBaseCon 2012. 5/22/12 12

Memory: JVM-specific metrics

• GC
• TIP: Correlate with “concurrent mode failure” or “promotion failed” in the
GC logs for GC pauses
• Heap usage
• TIP: Do not want this to max out – add more memory or assign more heap
to your RS
• Number of logs categorized by level
• TIP: Be careful of any WARN or ERROR logs – helpful first indicator to look
more deeply in logs
• Threads blocked/running

HBaseCon 2012. 5/22/12 13

Latency: FS read

• h.r.fsReadLatency_avg_time – average time per sequential read in ms
• h.r.fsReadLatency_num_ops – number of sequential reads
• h.r.fsPreadLatency_avg_time – average time per positional read in ms
• h.r.fsPreadLatency_num_ops – number of positional reads
• Various histograms based on this – 75th, 95th percentile, mean, standard
deviation, etc., in nanoseconds
• TIP: Spikes in h.r.fsReadLatency_avg_time/h.r.fsPreadLatency_avg_time
can indicate HDFS/disk/network problems
HBaseCon 2012. 5/22/12 14

Latency: FS write/sync
• h.r.fsWriteLatency_avg_time – average time per HLog edit write in ms
• h.r.fsWriteSize_avg_time – average size per HLog edit write in bytes
• h.r.fsWriteLatency_num_ops, h.r.fsWriteSize_num_ops – number of HLog
edit writes
• h.r.fsSyncLatency_avg_time – average time to sync HLogs to the filesystem
in ms
• h.r.fsSyncLatency_num_ops – number of HLog syncs
• h.r.slowHLogAppendCount – number of HLog edits that took longer than 1
second
TIP: Search for a WARN-level log message with the text “appending an edit to hlog”;
use that time to spelunk through logs to see what else is going on
• Various histograms based on this – 75th, 95th percentile, mean, standard
deviation, etc., in nanoseconds
• TIP: Spikes in write or sync latencies can also be due to HDFS/bad
disks/bad network

HBaseCon 2012. 5/22/12 15

Latency, region: per-operation-type, per-region
• From regionserver metrics: h.r.*RequestLatency - latency histograms for
various client operations in nanoseconds
• Also general RPC metrics for all requests from clients
• rpc.metrics.RPCQueueTime is time it takes from receipt of the RPC until it starts
being processed
• rpc.metrics.RPCProcessingTime is time it takes from start to end of processing
• TIP: These are some of the first places to look if your client seems to be running
slow
• h.r.readRequestsCount – requests for gets, scanners. Not # of rows.
• h.r.writeRequestsCount – requests for puts, deletes, etc. Not # of rows.
• hbase.RegionServerDynamicStatistics.* contains some metrics on a per-
region and per-column-family basis
• Subset of block cache, compaction, flush, store/storefile sizes, etc.
• TIP: These metrics can be used to answer the questions:
• Is the problem only in one region or in all regions?
• Is the problem only with one operation type or all operations?

HBaseCon 2012. 5/22/12 16

Host-level metrics
• CPU:
• Load averages
• % idle/system/user/wio (waiting for block I/O)
• Running/total processes
• Correlate with “top”, “sar”, “mpstat” to figure out which process is
doing what.
• TIP: If a process is eating a lot of CPU, that is a clue to look at its logs
• Memory:
• swap_free/swap_total
• cached/free/shared
• TIP: Compare with logs, other metrics to help determine causes of
OOMs

HBaseCon 2012. 5/22/12 17

Host-level metrics
• Network:
• Bytes/packets sent/received
• TIP: Correlate with ZK timeouts from logs if RSes are going down
• TIP: Also look at Ethernet frame errors from ifconfig if you are
experiencing unexpected dips in network traffic
• Disk I/O:
• Read/write latencies
• Slow disks make HBase performance very bad, especially on .META.
• TIP: Check dmesg for SCSI errors
• disk space available
• swap usage
• TIP: Should be 0 – otherwise you will have timeouts and RSes going
down. Add RAM and/or buy more boxes.
HBaseCon 2012. 5/22/12 18

Takeaways

• Know what you are looking for
• Know what metrics are available, what they mean
• Collect a lot of different metrics, but don’t drown in them
• Metrics are not all you need
• Correlate metrics with logs, workload
• Metrics are another tool in your toolbox. You still have to do
the work to monitor/debug/tune.

HBaseCon 2012. 5/22/12 19

Takeaways
• Take a baseline of your
system in steady-state
• For later comparison if things
go bad
• Try to make this as apples-to-
apples as possible, e.g. same
hardware/workload/config
• Spikes or dips from
baseline can indicate
problems
• But also depends on which
metric and if you can explain
it (e.g. increased workload)

HBaseCon 2012. 5/22/12 20

Thank you
(Especially to Jon Hsieh and Todd Lipcon
for their helpful reviews and suggestions)

Cloudera is hiring
E-mail me: dsw@cloudera.com

HBaseCon 2012. 5/22/12 21

HBaseCon 2012 | Base Metrics: What They Mean to You - Cloudera

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to HBaseCon 2012 | Base Metrics: What They Mean to You - Cloudera (20)

More from Cloudera, Inc. (20)

Recently uploaded (20)

HBaseCon 2012 | Base Metrics: What They Mean to You - Cloudera

Editor's Notes